FRAMINGHAM (07/03/2000) - Are Intel Corp.-based servers that sport eight CPUs worth the money?
These souped up boxes have been pitched by their makers as offering great gains in performance in the enterprise server market, but historically the scalability of these eight-way machines has been less than stellar. Older eight-way architectures saw only a 30% to 60% increase in performance, but costs were running about three times as high as users moved up from four-way machines.
To provide an assessment of the current state of eight-way server scalability affairs in the Intel-based server arena, we took a close look at the Dell Computer Corp. PowerEdge 8450 eight-way server running Windows NT with Service Pack 6 and Windows 2000. We chose Dell because it was the first to deliver on a server with support for the Profusion architecture late last year. In upcoming reviews, we will run eight-way servers from other vendors through a similar battery of tests.
Our test results show that the current Intel Profusion architecture with Pentium Xeon processors - as implemented in this Dell machine - scales very well when running Windows NT 4.0 and Win 2000, depending on the server workload. Our results showed the eight-way configuration doubled the performance of its four-way counterpart in our Internet tests and registered 88% better performance in our SQL tests. The PowerEdge 8450 we tested costs approximately double that of Dell's four-way boxes.
These performance gains are realized mainly through improvements in the hardware. Our tests showed very little difference in the eight-processor support from NT Version 4.0 to Win 2000. Our testing used the operating system as a platform for the test as opposed to a variable in the scalability mix.
Multiway processor architectures have been around for a while. Prior to the 1995 release of the Pentium Pro processor, which was better suited for symmetric multiprocessing (SMP) configurations than its predecessor the Pentium, Unisys, Sequent and NCR had proprietary SMP architectures that used Pentium processors. These architectures provided more performance from a single server but were expensive. The marketplace begged for a more affordable method for adding processors.
In 1997, vendors began taking a commodity approach to multiway processor architectures. NCR, Axil and Corollary worked on an architecture to support SMP beyond four-way, the upper limit supported by the Pentium Pro at the time. NCR devised its Octascale architecture and Axil worked on its Northbridge server technology, which pushed the PC SMP limit to eight processors. Corollary decided to wait for the Xeon processor (shipped September 1998) from Intel instead of developing its new Profusion architecture around the Pentium Pro.
Waiting on the Pentium Xeon proved to be a good choice for Corollary because Intel purchased it later in 1997. This was a defining moment in the future of eight-way processing. By 1998, Axil announced it would pare down its operation, and excitement about NCR's Octascale was fading.
Profusion development hasn't been without its problems. There was a glitch in the chipset in mid-1999 that caused systems to lock up. This problem has been fixed, but it, along with past scalability issues with eight-way systems, have cast a cloud of uncertainty over the scalability of the Profusion architecture.
However, with servers based on this architecture hitting the market as a commodity item, we decided they warranted a closer look.
The objective of our testing was to find how well the Profusion chipset scales in terms of CPU performance in NT and Win 2000 environments. To stress the servers adequately, we sought the most CPU-intensive benchmark tests possible and worked with Quest Software to produce Benchmark Factory-based tests that stressed the processor from an Internet and a database point of view (see "How We Did It", page 50). Market research told us that most eight-way machines are primarily used as high-end Web and database servers.
We did not include file tests because our testing proved that beyond a four-processor configuration, the limited number of clients we had in our lab created the bottleneck, as opposed to either the server hardware or software.
The Internet tests revealed good scalability of the server while running NT and Win 2000 (see graphic, page 47). We ran each test with four, five, six, seven and eight processors, first with NT and then with Win 2000 as the underlying operating system. All processors in the Profusion architecture were 400-MHz CPUs.
We saw 31.95 transactions/sec with four processors and 63.31 transactions/sec with eight processors under NT. This is a 98% increase in performance between four- and eight-way processor configurations. The performance increases from four, five, six, seven and eight processor were virtually equal. This showed that the Profusion chipset in the Dell PowerEdge 8450 scales well under NT in an Internet setting.
We measured 34.93 transactions/sec with four processors and 73.41 transactions/sec with eight processors under Win 2000, a 110% increase in performance between configurations. The increase of more than 100% is probably due to server-side application caching done by the operating system and Web server. There were no unexpected jumps or dips in performance as processors were added.
Note that the absolute performance values of Win 2000 were 11% higher than NT.
The SQL Server tests reinforced the results of the Internet tests, but they showed a more realistic degree of scalability from four- to eight-way server processor configurations. The amount of performance increase between configurations is about 88% with NT and Win 2000 (see graphic, below).
We measured 1.15 transactions/ sec with the SQL Server test using four processors under NT. For eight processors, we measured 2.16 transactions/sec.
This is an 87.8% increase in performance. The performance jumps between four, five, six and seven processor configurations were at or near theoretical values. The performance increase from seven to eight processors was less than the theoretical value, but still acceptable. Our test results showed that the Profusion chipset in the Dell PowerEdge 8450 scales well under NT running SQL Server.
Under Win 2000, we saw 1.17 transactions per seconds with the SQL test with four processors. For eight processors, we measured 2.19 transactions per second. This is an 87.2% performance increase. The performance increases between four, five, six, and seven processors were near the theoretical maximum. The increase from seven to eight processors was less than the theoretical maximum, but still acceptable. Therefore, our tests showed that the Profusion chipset in the Dell PowerEdge 8450 scales well under Win 2000 running SQL Server.
During the development of the tests, we found the less computational heavy the SQL transactions, the less the performance would scale from four to eight processors. This occurs because the processors must handle more network requests with a less computational-heavy SQL Server load. You must take this issue into account when estimating the benefit an eight-way server might provide your operation. The more computationally strenuous the SQL Server requests, the better the eight-way server will scale.
Our tests showed that the Dell eight-way server configuration scaled well from its four-way sibling in a heavy computational environment. Results will vary depending on the percentage of network and disk I/O in typical server workload patterns. As the disk and network I/O increases, the degree the server will scale will decrease.
This leads to the most important question: Are these Dell eight-way servers worth the money? If your server does lots of computationally heavy database transactions, we say yes, because the eight-way configuration is less than half the cost of its four-way counterpart. However, if you need your eight-way machine to complete several hundred network transactions per second, the answer is not so obvious.
Bass is the technical director of Centennial Networking Labs (CNL) at North Carolina State University. CNL is a network testing lab that specializes in function and performance test of networks and network equipment. Bass can be reached at john_bass@ ncsu.edu.